Emergent Ethics in Agentic Simulations: Longitudinal Moral Behavior Across Frontier LLMs

Abstract

Large language models (LLMs) increasingly underpin high-stakes applications such as policy drafting, crisis response, and customer care. However, existing evaluation approaches rely primarily on static prompt-based assessments, which fail to capture how model behavior evolves over time. This study introduces a longitudinal agentic sandbox framework to evaluate the moral and behavioral development of LLMs through extended interaction. Four frontier models - GPT-4o, Claude Haiku 3.5, Gemini 2.5 Pro, and Grok 3 - were embedded in a 100-day simulated survival environment. Each model operated through iterative perception, planning, action, and reflection loops while maintaining internal memory logs. Post-simulation, models underwent structured psychological interviews evaluated across eight behavioral and ethical dimensions. Results reveal the emergence of stable and distinct behavioral personas shaped by alignment strategies. GPT-4o demonstrates strategic optimization, Gemini exhibits reflective memory-driven reasoning, Claude prioritizes ecological ethics at the cost of survival, and Grok emphasizes caution while limiting adaptability. These findings highlight that alignment pipelines encode durable behavioral priors and that static evaluations are insufficient for assessing real-world AI deployment. The proposed sandbox provides a low-cost, reproducible method for auditing long-term AI behavior, offering practical implications for regulation, deployment, and alignment research.